RTB rips black box part 2:algorithm meets World

Part 0 describes the winning price of RTB rising steeply in the wee hours of the morning. We also introduced a Pace system, and if all the DSPs of the system were used, the problem of steep rise would disappear. the system in part 0 contains an implicit hypothesis that any two requests are considered to be the same, ignoring other factors, such as request time. The competitive price described in part 1 will change over time, but this is ignored by the system in part 0 .

Some quick-thinking readers are aware of the problem, and they have commented that it is not a good time to bid if they know that some of the time's winning prices are high.

This article describes how we calculate the profit on a per-request basis and explains how a Pacer simultaneously handles changes in competitive prices and display quality, and does not require hard-coded time to bid. To achieve the above target only need to pacing System Update control variable period is much longer, such as one day, and the control variable as a "minimum acceptable expected profit" threshold value. Is the comparison between the average competitive price and the average expected profit over time, the graph collects multi-day data, notice can see the early morning price steep rise, profit decline:

A well-adjusted, infrequently changing pacing system will eventually appear as a horizontal line in the graph, and we only bid for requests that expect a profit exceeding the threshold. The core idea of this approach is that the curve in the graph represents a distributed average of a wide and skewed distribution, so even if the average price goes up and the expected profit falls, it does not mean that there is no good request, it simply means that there are fewer good requests. By setting a fixed, infrequently updated threshold, we can bid at any time of the day, but may consume less in the wee hours and consume more before the wee hours.

This can be said to take advantage of other bidders ' suboptimal behavior, because some bidders irrational in the early morning into the auction, so that the market price steep rise.

Our algorithm is the use of scraping cream, only the best bid for the request. We have reservations during the day, and when the other bidder's budget is exhausted, the demand for the request drops, which leads to a drop in the market price and, at the same time, a rise in the expected profit, we will increase the amount of competition.

Navigating Around Publisher price Floors

We've introduced our algorithmic strategy for other bidders, and we'll cover policies for publishers. How does this system deal with the publisher's reserve price? The bottom price is the rate at which the publisher is willing to sell. In fact, it can be understood that the publisher is also involved in this ad bit auction. If no bidder exceeds this reserve price, there is no bidder to get the exposure. If more than one bidder is above the reserve price, then the reserve is not affected, and the winner is still paying the second place. If only one bid is higher than the reserve price, then the Publisher will get some more profit.

First Price	Second Price	Reserve price	Clearing price
5	3	N/A	3
5	3	2	3
5	3	4	4
5	3	6	No transaction

This is very reasonable at the micro level, if we win the bid, we are happy to receive a higher deduction fee: Because our bid is the value we believe, we pay the second place bid, the difference between the two is our profit. Publishers are more likely to be charged for our bids, because they are motivated to keep the bottom line up so they can get more profit. Our algorithm uses the method of dealing with bidders to deal with the publisher's elevation reserve. At any one time, thePacer control variable tells us the minimum expected profit threshold. If publishers raise their bottom line, the expected profit for their display opportunities will fall. If it falls below the threshold, we will not bid on this request, but will not accept it to improve the bottom line. This is what we mean by viewing the reserve price. The setting of the control variable is determined by the forecast and is based on the profit space of the exchange as a whole. So any publisher who raises the reserve price may lose our bid because we can get exposure from other publishers. The bottom line may be set lower than our bid, but may be higher than the second bid, so we will give up a portion of the profit to the publisher. This is the meaning of exchange : Market mechanisms use competition to make profit margins bigger.

It is important to note that at the macro level, publishers do not have to sell their traffic all over the RTB : They can sell it at a higher price by way of contract advertising. In terms of advertising terminology, contract advertising traffic is "high-quality traffic", the other traffic is "residual traffic." Publishers are motivated to set the reserve price higher because advertisers may get lower-priced advertising results, such as RTB . Advertising terminology is a "cross-channel conflict" because the remaining traffic can hurt the benefits of high-quality traffic. If the publishers set the reserve price for this reason, they might set it higher than our bid, rather than trying to squeeze the profit between first and second place. The result is that the expected profit is zero, and as a DSP We do not need to bid for the non-profit request. Mike on Ads has an article on this question that explains why it's not a good idea for publishers to do so.

RTB rips black box part 2:algorithm meets World